The Impact of Tutorials on Games of Varying Complexity

 
CONTINUE READING
The Impact of Tutorials on Games of Varying Complexity
The Impact of Tutorials on Games of Varying Complexity

  Erik Andersen, Eleanor O’Rourke, Yun-En Liu, Richard Snider, Jeff Lowdermilk, David Truong,
                                  Seth Cooper, and Zoran Popović
                                      Center for Game Science
              Department of Computer Science & Engineering, University of Washington
           {eland,eorourke,yunliu,mrsoviet,jeff,djtruong,scooper,zoran}@cs.washington.edu

ABSTRACT                                                                             Learnability is widely accepted as a central component of us-
One of the key challenges of video game design is teaching                           ability, and it has been studied in the field of Human Com-
new players how to play. Although game developers fre-                               puter Interaction for decades [11, 20, 23]. In this paper,
quently use tutorials to teach game mechanics, little is known                       we draw tutorial ideas from the HCI literature and present a
about how tutorials affect game learnability and player en-                          large-scale comparative study of their effectiveness in video
gagement. Seeking to estimate this value, we implemented                             game tutorials. Although many factors influence tutorial de-
eight tutorial designs in three video games of varying com-                          sign, we chose to focus on four specific tutorial characteris-
plexity and evaluated their effects on player engagement and                         tics: the presence of tutorials, the context-sensitivity of tuto-
retention. The results of our multivariate study of over 45,000                      rial instructions, the freedom given to users during the tuto-
players show that the usefulness of tutorials depends greatly                        rial, and the availability of additional help on demand. We
on game complexity. Although tutorials increased play time                           examined the importance of these characteristics by conduct-
by as much as 29% in the most complex game, they did not                             ing a multivariate experiment across three games that we de-
significantly improve player engagement in the two simpler                           veloped: Refraction, Hello Worlds, and Foldit. These games
games. Our results suggest that investment in tutorials may                          belong to different genres and vary in complexity.
not be justified for games with mechanics that can be discov-
                                                                                     We present results gathered from over 45,000 players show-
ered through experimentation.
                                                                                     ing that the value of tutorials depends greatly on the com-
                                                                                     plexity of the game. In Foldit, the most complex and least
Author Keywords                                                                      conventional game we studied, tutorials increased play time
games; analytics; tutorials; multivariate testing                                    by as much as 29% and player progress by as much as 75%.
                                                                                     However, we found that tutorials had a surprisingly negligible
ACM Classification Keywords                                                          effect on player engagement in Hello Worlds and Refraction,
H.5.0 Information interfaces and presentation: General                               which are less complex and more similar to other games in
                                                                                     their respective genres. Giving tutorial instructions as closely
INTRODUCTION                                                                         as possible to when they were needed, rather than out of con-
Teaching new players how to play a game is challenging but                           text in an up-front manual, increased play time by 16% and
crucial for engaging and retaining players. As a result, game                        progress by 40% in Foldit but had no effect in the other two
designers frequently utilize tutorials to aid the learning pro-                      games. We found no evidence to support the claim that re-
cess. Since tutorials are typically the first part of a game that                    stricting player freedom in order to focus the player’s atten-
new players encounter, effective tutorial design is important                        tion on a target concept improves learnability. Providing help
for retaining new players. Although modern games employ                              on-demand improved player engagement in Foldit, but had
a wide varety of tutorial styles, including hints, help buttons,                     no effect in Hello Worlds and even negative effects in Refrac-
manuals, and interactive training challenges, the relative ef-                       tion. Our results suggest that investment in tutorials may not
fectiveness of these styles is not well understood. As a result,                     be justified for games that can be learned through experimen-
designers must rely on intuition, personal experience, existing                      tation, and that the players of these games, at least, seem to
examples, and extensive user testing when designing tutori-                          learn more from experimentation than from reading text.
als. A deeper understanding of how tutorial design decisions
impact player engagement and retention would help game de-
velopers create more learnable interfaces and spend valuable                         Tutorial presence
resources more wisely.                                                               The first feature we considered was tutorial presence:
                                                                                     whether the game provides a tutorial to its players or not.
                                                                                     Although some successful games provide no instructions to
                                                                                     players, such as Tetris (Pajitnov 1984), Pac-Man (Namco
Permission to make digital or hard copies of all or part of this work for            1980), and Super Mario Brothers (Nintendo 1985), the use
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
                                                                                     of tutorials has become very common. Therefore, we wanted
bear this notice and the full citation on the first page. To copy otherwise, or      to examine the impact of tutorial presence.
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.                                                             Hypothesis 1: Games with tutorials will exhibit better player
CHI’12, May 5–10, 2012, Austin, Texas, USA.                                          engagement and retention than games without tutorials.
Copyright 2012 ACM 978-1-4503-1015-4/12/05...$10.00.
The Impact of Tutorials on Games of Varying Complexity
We examined this hypothesis by comparing versions of each         In this study, we compare tutorials that present information
game with and without tutorials.                                  out of context in an up-front manual and tutorials that present
                                                                  information in context just before the player needs to use a
                                                                  particular concept.
Context-sensitivity
Another feature we considered was the context-sensitivity of      Freedom
tutorials. Existing tutorials can be divided into two cate-       We also considered tutorial freedom, or the degree to which
gories: those that provide contextually relevant suggestions      tutorials force users to perform the mechanics being de-
from within the application interface and those that provide      scribed. While requiring users to practice specific game me-
documentation outside of the application context. HCI re-         chanics could improve tutorial effectiveness, it could also
searchers have studied both types of tutorials extensively;       frustrate players. As a result, it is not clear if and when re-
however, the effect of context-sensitivity on tutorial success    ducing player freedom improves engagement.
is unclear, and the existing literature does not show a clear
preference for either type of information.                        Game designers have often cited the importance of allow-
                                                                  ing users to experiment while learning new concepts [7, 22].
Historically, tutorials have been provided through primarily      James Gee argues that such experimentation is most effective
textual documentation that is accessed outside the applica-       when it occurs in a safe environment where functionality is
tion context. While paper manuals and online documentation        reduced and mistakes are not punished [7]. Software learn-
require minimal resources to generate, they present a number      ability researchers Kelleher and Pausch argue that restrict-
of challenges for users, who struggle to keep track of instruc-   ing user freedom improves tutorial performance. They found
tions when switching between the tutorial and the application     that an interactive tutorial for the Alice computer program-
[1, 15, 17]. As a result, many researchers have worked on         ming environment that restricts user freedom, called Stencils,
improving external software documentation by incorporating        helped users finish the tutorial sequence more quickly and
screenshots, graphics, animations, and annotations into tex-      with fewer errors than a paper-based tutorial with the same
tual content [9, 12, 13, 21].                                     content [14]. How different types of interfaces are affected
Despite these improvements, many researchers argue that           by reduced player freedom, however, is currently unknown.
contextually relevant tutorials have greater potential for im-    Many video games provide players with complete freedom
proving application learnability [4, 14, 10]. This idea is sup-   during the tutorial. Braid (Number None 2008), for instance,
ported by situated learning, a popular learning model in ed-      provides the player with background suggestions for keys to
ucation based around teaching concepts in the same context        press during gameplay, but does not force the player to per-
in which they will be applied [18]. Game expert James Gee         form the suggested action. At the same time, the use of
also highlights the importance of teaching new mechanics just     more restrictive stenciling strategies is also popular, and can
before the player needs to use them, rather than presenting       be found in PopCap’s tower defense game Plants vs. Zom-
them out of context [7]. Tooltips, which provide brief textual    bies (PopCap 2009), the simulation game SimCity 4 (Maxis
help when the user hovers over an interface component, are        2003), and Facebook game Cityville (Zynga 2010). We in-
one of the most successful forms of contextual help [6]. On       cluded tutorials of both types in order to determine the value
the other hand, context-sensitive help can be more frustrating    of restricting player freedom.
than helpful, as shown by Microsoft’s Office Assistant [24].
                                                                  Hypothesis 3: Tutorials that restrict player freedom improve
Commercial video games have used both context-sensitive           engagement and retention.
and context-insensitive tutorials successfully. Up-front man-
uals, which present game mechanics out-of-context at the          In this study, we compare tutorials that restrict player free-
very beginning of a game, can be seen in the popular on-          dom using the stenciling technique proposed by Kelleher and
line Flash games Cursed Treasure1 and Epic Battle Fantasy         Pausch [14] and those that give players complete freedom to
32 . Other games present instructions in-context, just as the     ignore the tutorial content.
player needs to use them. This strategy can be seen in the
highly-rated Flash game In the Company of Myself3 and the         Availability of help
popular console game Portal (Valve Corporation). To learn         Finally, we considered the availability of help. Many appli-
more about the value of context-sensitivity in game tutorials,    cations provide access to documentation on-demand through
we included both styles of presentation.                          help buttons, but it is unclear whether this type of informa-
                                                                  tion has a strong effect on interface learnability. Although it
Hypothesis 2: Tutorials that present instructions in context      seems likely that providing access to on-demand help would
will be more effective than tutorials that present information    positively impact learnability, it is not known how much users
out of context.                                                   take advantage of on-demand information when it is avail-
                                                                  able.
1
  http://www.kongregate.com/games/IriySoft/                       James Gee argues that manuals are most effective when they
cursed-treasure-dont-touch-my-gems
2
  http://www.kongregate.com/games/kupo707/                        are accessed on-demand, after the user has become familiar
epic-battle-fantasy-3                                             with the basic interface elements and has developed specific
3
  http://www.kongregate.com/games/2DArray/                        questions [7]. Commercial software applications rely heav-
the-company-of-myself                                             ily on the availability of on-demand documentation, and help
The Impact of Tutorials on Games of Varying Complexity
buttons can be found in many ubiquitous products such as Mi-
crosoft Word (Microsoft Inc. 2010) and iTunes (Apple Inc.
2001). On-demand tutorials are also common in games. For
example, help buttons are available in the Flash game Epic
Battle Fantasy 3 and the real time strategy game Starcraft 2
(Blizzard 2010). To determine the impact of on-demand help
in games, we included tutorials with and without this feature.
Hypothesis 4: Having on-demand access to help improves
player retention.
In this study, we compare games that provide access to tu-
torial manuals on-demand through a help button with games
that do not, both with and without other types of tutorial in-
structions.
                                                                    Figure 1. A level of Refraction. The goal is to use the pieces on the
                                                                    right to split lasers into fractional pieces and redirect them to satisfy the
STUDY METHODOLOGY                                                   target spaceships. The user can pick up and put down pieces by clicking
Over the past decade, researchers have begun to use the Inter-      on them. The pieces are designed to be as intuitive as possible. The grid
                                                                    interface, pipe-flow game mechanics, and spatial reasoning puzzles are
net to collect large quantities of data on player behavior. This    similar to many other puzzle games.
has given rise to a data-driven evaluation methodology that
has been widely used by companies like Amazon, Microsoft,
and Google to support interface design decisions [16][19][8].       Although Refraction was originally designed to teach math
Such evaluation models frequently use simple A/B testing, in        concepts, the game has found success with an adult audi-
which a baseline sample is compared to a variety of single-         ence on the popular Flash website Kongregate, where it has
variable test samples. This methodology has been used to            been played over 480,000 times since its release in Septem-
evaluate games as well; Andersen et al. used A/B testing to         ber 2010. The use of fractions as a game mechanic is rela-
show that gameplay affects average play time more than an-          tively uncommon, but similar spacial reasoning puzzles and
imations and audio [2], and that secondary game objectives          pipe flow mechanics appear in many other games on Kongre-
are most effective when they support the primary objectives         gate and elsewhere. The artwork for the laser sources, target
[3].                                                                spaceships, asteroids, and laser manipulators was designed to
                                                                    utilize affordances to make the functionality of each piece as
In this study, however, we wanted to compare a large vari-          clear as possible. Refraction is freely available and can be
ety of tutorial designs defined by multiple variables. We also      played by anyone with a web browser.
had a limited understanding of which combinations of tuto-
rial characteristics would be most effective. As a result, we
chose to perform a multivariate experiment considering four
independent variables and a total of eight conditions, rather
than using a simple A/B test.

Our games
We performed tutorial experiments on three games developed
by our research group: Refraction, Hello Worlds, and Foldit.
All three games contain puzzle elements, but otherwise dif-
fer greatly in their format, complexity, and player bases. We
consider two of the games, Refraction and Hello Worlds, to
be “casual” games. We define “casual” games as those that
rely on familiar game mechanics, take only a few hours to
complete, and do not require a download.
Refraction, which involves solving spacial puzzles by split-        Figure 2. A level of Hello Worlds, a puzzle-platformer game. The game
ting lasers into fractional amounts, was originally designed        interface is similar to thousands of other games in the platformer genre,
as a math-learning game. Each level is played on a grid that        except that the player exists in multiple worlds at the same time. As a
                                                                    result, the basic game mechanics for navigating using keyboard input
contains laser sources, target spaceships, and asteroids, as        may be familiar to players. The player advances through each level by
shown in Figure 1. Each target spaceship requires a fractional      opening and closing doors, a common goal mechanic.
amount of laser power, indicated by a yellow number on the
ship. The player can satisfy the targets by placing pieces that     Hello Worlds is a puzzle game that offers a simple twist on
change the laser direction and pieces that split the laser into     the standard platformer genre. Platformers typically involve
two or three equal parts. All targets must be correctly satisfied   navigating a character through a two-dimensional world by
at the same time to win.                                            running, jumping, and avoiding obstacles. In Hello Worlds,
The Impact of Tutorials on Games of Varying Complexity
the character exists in multiple worlds at the same time, and             advantage is that, in contrast to laboratory user studies, our
interacts with the obstacles in each world simultaneously, as             experiments are conducted “in the wild” where players do not
shown in Figure 2. The player must find a path through the                know that they are being observed. As a result, participants
worlds to reach various doors that open and close worlds or               are playing under their own motivation, and our findings re-
complete the level. While the puzzle element of Hello Worlds              flect their natural behavior. However, one important limita-
is uncommon, the basic game mechanics for navigating using                tion is that we have no interaction with our participants, and
keyboard input are used in many existing platformers. Hello               we cannot know what they are thinking or feeling. Therefore,
Worlds is available for free on the Flash game website Kon-               we must infer the effects of tutorials on player engagement
gregate, and has been played over 1,300,000 times since its               by measuring their behavior.
release in May 2010.
                                                                          We measure player engagement in three ways. First we count
                                                                          the number of unique levels each player completes. Sec-
                                                                          ond, we calculate the total length of time that they played
                                                                          the game. Since players occasionally idle for long periods,
                                                                          we aggregated moves the player made in 30-second intervals,
                                                                          removing periods of two or more consecutive idle intervals
                                                                          from the total play time. Finally, we measured the return rate,
                                                                          defined as the number of times players loaded the game page
                                                                          in Refraction and Hello Worlds, and as the number of times
                                                                          the game was restarted in Foldit. While these metrics are not
                                                                          precisely the same as “engagement” or “learnability,” we ar-
                                                                          gue that they are closely related, given that players are free
                                                                          to leave at any time and cannot advance in a game without
                                                                          learning its associated rules and strategies. If the purpose of
                                                                          tutorials is to teach game mechanics to the player, the ability
                                                                          to complete levels should be the ultimate test of whether the
                                                                          player has learned those mechanics.
Figure 3. A puzzle in Foldit, a game in which players manipulate pro-
tein models and try to pack the shape as tightly as possible. The game
                                                                          We modified each of the games to randomly assign new play-
is complex, has an interface with many tools and buttons, and offers a    ers to one of the eight experimental conditions. We focused
gameplay experience that is very different from other games. Therefore,   solely on new players for this experiment; veteran players
players cannot rely on prior knowledge of other games very much when      who were already familiar with the games were not included
trying to learn how to play Foldit.                                       in our analysis. In Refraction and Hello Worlds, we tracked
                                                                          players using the approach described in [2], storing progress
Foldit is a multiplayer, online game in which players com-                and experimental condition through the Flash cache. One
pete and collaborate to fold protein structures efficiently [5].          drawback of this method is that if players changed computers
Players can manipulate 3D protein structures, shown in Fig-               or deleted their Flash cache, they were treated as new players.
ure 3, by pulling on them, freezing pieces to prevent them                However, since the Flash cache is inconvenient to clear and
from moving, and launching optimizations that will compu-                 this action deletes all progress in the game, we considered
tationally improve the protein. Each folded protein is given              this risk to be small. For Foldit, we tracked players using
a score which is used to rank solutions. Foldit has a set of              their Foldit account and stored the condition number locally
offline tutorial puzzles designed to prepare players for online           on their machine. Players who logged in on multiple ma-
competitions, which we used for this study.                               chines were tracked by their account and discarded to prevent
Foldit differs from Refraction and Hello Worlds in several key            the inclusion of players who played multiple tutorial versions.
ways. First, players must download and install Foldit from a
website in order to play, and must create an account to partic-           STUDY DESIGN
ipate in online competitions. Second, the gameplay mechan-                We performed a multivariate experiment with four variables:
ics required to fold proteins are unique to this game, and as             presence, context-sensitivity, freedom, and availability of on-
a result it is unlikely that players can benefit from past gam-           demand help. We picked these variables to capture the most
ing experience when playing Foldit. Finally, it is much more              salient differences in tutorial implementations seen in suc-
complex than the two “casual” games and requires an under-                cessful games. We included a total of eight experimental con-
standing of some unconventional mechanics, such as placing                ditions, shown in Table 2, each representing a combination of
rubber bands to hold the protein together, or running a wide              values for the four variables. Table 1 shows a breakdown of
variety of optimizations. Since Foldit players are working to             the concepts that each game introduces and the order in which
solve complex, open scientific problems through gameplay,                 they are introduced. The following sections explain how the
we do not consider this a “casual” game.                                  four variables are implemented in each game.

Metrics                                                                   Tutorial presence
Collecting data on player behavior through large-scale anony-             We wanted to examine how the addition of tutorials affected
mous experiments has strengths and weaknesses. One key                    player behavior. Therefore, the first variable in our experi-
The Impact of Tutorials on Games of Varying Complexity
Hello Worlds                                       Refraction                                     Foldit
                     Concept            Level   Page                 Concept             Level   Page              Concept             Level   Page
              Character Movement (S)      1      1               Powering Ships            1      1                Clashes               1       1
                 Collecting Coins         1      2                  Benders (S)            1      1          Pulling Sidechains          1       1
                 Using Doors (S)          1      1              Multiple Benders           3       -        Rotating the Camera          2       2
                    Red Doors             1      3             Ship Input Direction        3      4                 Score                2       2
                 Overworld Doors         O       5                   Asteroids             4      1               Shake (S)              3       3
                    Star Types           O       6                     Coins               5      3       Pulling the Backbone (S)       4      4,5
                 Multiple Worlds          2      1                 Splitters (S)           6      5                 Undo                 4       8
                 Combo View (S)           2      4             Fractional Ships (S)        6      4                 Voids                5       6
                    Rewind (S)            3      4             Magnifying Glass (S)        7      6                 Reset                5       8
                   Green Doors            6      3                Combiners (S)           33      7              Wiggle (S)              6       7
                    Blue Doors           10      3                Expanders (S)           37      8           Hydrogen Bonds             7       9
                                                              Multiple Input Ships (S)    43      9           Wiggle Again (S)           7       7
                                                              Multiple Denominators       44      9          Rubber Bands (S)            8      10
                                                                                                           Translating the Camera        9       2
                                                                                                          Rubber Bands Again (S)         9      10
                                                                                                                 Freeze (S)             10      11
                                                                                                               Backbone Color           11      12
                                                                                                                 Rebuild (S)            11      12
                                                                                                                Hydrophobics            12      13
                                                                                                                  Exposeds              12       -
                                                                                                              Tweak Rotate (S)          13      14
                                                                                                               Tweak Shift (S)          14      15
                                                                                                           Tweak Rotate Again (S)       15      14
                                                                                                        Secondary Structure Mode (S)    16      16
Table 1. A breakdown of concepts taught in each game. The concepts are listed by the order in which they were taught in the context-sensitive condition.
Concepts were either taught in a specific level as part of a context-sensitive tutorial or in the manual used by the context-insensitive tutorial and the
on-demand help. Concepts marked with an “(S)” were stenciled during the blocking conditions. Foldit is more complex than the other two games and
required many more tutorials.

      Condition       Presence     Context       Freedom           Help
   Insensitive+Help    Present    Insensitive   Nonblocking      On-demand
                                                                                        In Refraction, as shown in Figure 4(a), the instructions ap-
      Insensitive      Present    Insensitive   Nonblocking        None                 peared in a notification bar at the top of the main grid. Re-
    Sensitive+Help     Present     Sensitive    Nonblocking      On-demand              fraction also contained “just-in-time” messages that would
       Sensitive       Present     Sensitive    Nonblocking        None
    Blocking+Help      Present     Sensitive     Blocking        On-demand              activate when the player tried to do something that was not
       Blocking        Present     Sensitive     Blocking          None                 allowed. For example, when a player powered a target space-
      Help Only         None         N/A           N/A           On-demand              ship with less than the required amount, a message saying
     No Tutorials       None         N/A           N/A             None
                                                                                        “too little power” appeared over the spaceship. Each of these
Table 2. The eight tutorial conditions that we implemented for our mul-
tivariate experiment, based on four independent variables: presence,                    messages appeared every time the spaceship was underpow-
context-sensitivity, freedom, and availability of help. These conditions                ered, but player could click “don’t show again,” which would
were designed to represent a wide range of existing tutorial designs and                cause that particular message to stop appearing.
evaluate multiple hypotheses simultaneously.
                                                                                        Foldit tutorial messages appeared in pop-up boxes as shown
                                                                                        in Figure 4(b). In Refraction and Foldit, longer context-
                                                                                        sensitive messages were sometimes displayed across multiple
ment was presence, resulting in two conditions: tutorials and
                                                                                        pages, which the user could advance by clicking on the “next”
no tutorials. For conditions in which there were no tutori-
                                                                                        or “Tell me more...” button.
als, the game explained nothing on its own and players were
forced to either experiment or seek help outside the game.                              In Hello Worlds, shown in Figure 4(c), help icons were scat-
The conditions with tutorials are further broken into more                              tered throughout the game to give the player information
categories in the following sections.                                                   when appropriate. These help icons were activated whenever
                                                                                        the player stood over them, as shown in 4(c). When activated,
                                                                                        the text relating to that help icon would appear in the space
Context-sensitivity                                                                     above the icon, allowing the user to easily read or ignore it.
To evaluate the effect of tutorial context-sensitivity on en-                           If the player’s avatar then moved off of the help icon, the
gagement and player retention, we introduced an additional                              text would immediately disappear. If the player moved their
variable, context-sensitivity. For conditions in which tutorials                        avatar back over the help icon, the text would reappear.
were present, we organized the tutorials from each game into
two versions: context-sensitive and context-insensitive.                                For the context-insensitive version, we grouped the concepts
                                                                                        of each game into multi-page manuals by topic and order
For the context-sensitive version, we tried to introduce each                           of introduction. Since each game requires a different num-
concept as closely as possible to when the player needed to                             ber of concepts, the length of the context-insensitive manu-
use it. Therefore, for each of the concepts in Table 1, we                              als varied across the games: 9 pages for Refraction, 16 for
added a short message explaining that concept to the first level                        Foldit, and 6 for Hello Worlds. The manuals for Refraction,
in the game requiring knowledge of it. Some tutorials also                              Foldit, and Hello Worlds are shown in Figures 5(a), 5(b), and
contained contextual images to reinforce the text. Each mes-                            5(c), respectively. The manual consisted of a brief textual
sage was controlled for size so that it could be easily ignored.                        description of each of the game’s concepts accompanied by
Additionally, the player could dismiss any tutorial message.
The Impact of Tutorials on Games of Varying Complexity
(a)                                               (b)                                                (c)
Figure 4. Screenshots of the context-sensitive tutorials. For Refraction, shown in Figure 4(a), a window appeared at the top of the screen containing
text, pictures, an OK button to continue, and a back button to return to a previous message. Foldit’s tutorial, shown in Figure 4(b), included a small box
containing tutorial text, which related to the protein at hand, and a “Tell me more...” button to go onto the next message. For Hello Worlds, shown in
Figure 4(c), blue help icons could be found throughout the game. Tutorial text appeared above an icon whenever the player’s avatar overlapped with it.

                        (a)                                               (b)                                                (c)
Figure 5. Screenshots of the context-insensitive tutorials. All three games featured a large window containing text, pictures, page numbers, “Next” and
“Previous” buttons, and a “Close” button to help the player navigate and understand the concepts of the game. The layout of this window varied from
game to game. Figure 5(a) shows the layout for Refraction, Figure 5(b) shows the layout for Foldit, and Figure 5(c) shows the layout for Hello Worlds.

relevant pictures associated with that concept. These manu-                     concepts that forced the player to use the new tool, blocking
als appeared before the player could interact with any levels,                  forward progress in the game until the player performed the
forcing them to interact with the manual in some way before                     desired action. After completing the stencil, freedom to inter-
playing the game. The player could navigate between pages                       act with the full user interface was restored. We implemented
with the “next” and “back” buttons and gauge their progress                     stencils slightly differently in each game due to their different
using page numbers present on each page. The player could                       interaction styles.
also close the manual at any time by clicking the “close” but-
                                                                                In Refraction, when the player encountered a stencil, the
ton.
                                                                                game drew the player’s attention to the target interface object
Due to their length, we broke the Refraction and Foldit tu-                     by making the entire window gray except for that object, dis-
torials into logical “chapters.” As the player advanced, new                    playing the instructions on a notification bar at the top of the
chapters became available. At each new stage, the context-                      screen, adding an animated yellow arrow that pointed to the
insensitive tutorial reappeared showing the first page in the                   object, and ignoring all user input directed at anything other
new chapter, and the player could navigate to previous chap-                    than the object. This implementation is shown in Figure 6(a).
ters if desired. Refraction had three chapters, displayed at the                If a level had both context-sensitive tutorial messages and a
beginnings of worlds one, five and six respectively. Foldit had                 stencil, then the game forced the player to click somewhere
four chapters, displayed every four levels.                                     on the screen to advance each tutorial message before forcing
                                                                                the player to complete the stencil instructions.
Freedom
To evaluate the effect of tutorial freedom, we divided the                      In Foldit, the game displayed a message telling the player
context-sensitive condition described above into two condi-                     what to do, as shown in Figure 6(b). The screen was not
tions: blocking and nonblocking. To create the blocking con-                    grayed out, and the user could still manipulate the camera
dition, we identified several concepts from each game that                      and interact with game’s user interface controls. However,
involved using a new tool, a new interface object, or a new                     the game prevented the player from manipulating the protein
keyboard key. We then added a “stencil” for each of these                       in any other way until they performed the desired action. In
The Impact of Tutorials on Games of Varying Complexity
(a)                                                (b)                                                (c)
Figure 6. Screenshots of the blocking tutorials. A blocking tutorial prevented the player from continuing until they completed a specified action. For
Refraction, shown in Figure 6(a), the tutorial grayed out the interface except for a window containing tutorial text, an object for the player to interact
with, and an arrow pointing at that object. Foldit, shown in Figure 6(b), used a small text box to tell the player what action they needed to perform.
Hello Worlds, shown in Figure 6(c), grayed out the screen except for character, the tutorial text, and the interface buttons whenever the blocking tutorial
was active.

some levels the player needed to advance through preceding                       DATA ANALYSIS AND RESULTS
context-sensitive tutorial messages to reach the stencil.                        We collected data from Refraction and Foldit for approxi-
                                                                                 mately two weeks, accumulating 13,158 Refraction players
In Hello Worlds, as in Refraction, a gray window was drawn
                                                                                 and 9,743 Foldit players. During this period, Kongregate
over the entire game except for the player’s avatar, the mes-
                                                                                 added a new badge to Refraction, pushing the game to the
sage text, and the user interface buttons. Since the user pri-
                                                                                 top of the “new badges list” and attracting many new players.
marily interacts with Hello Worlds through the keyboard, the
                                                                                 Kongregate also featured Hello Worlds on the front page of
displayed message associated with each stencil told the user
                                                                                 its website as a “badge of the day,” allowing us to collect data
what key(s) to press to continue, as shown in Figure 6(c). All
                                                                                 from 22,417 Hello Worlds players over a two-day period.
other keys were ignored until the user pressed one of the valid
keys. The user was allowed to interact with the interface but-                   Table 3 shows the results of our experiment. Our mea-
tons, such as “quit,” but could not make any progress on the                     surements of levels completed and time played were not
level until completing the instructions in the stenciled mes-                    normally distributed, so we used on a non-parametric test,
sage.                                                                            the Wilcoxon/Kruskal-Wallis 2-sample test, to analyze lev-
                                                                                 els completed and time played for large-scale effects. The
Availability of help
                                                                                 Z-value reported for this test is a standard normal random
                                                                                 variable, a scaled version of the distribution of relative rank-
To evaluate whether players would look for help when they
                                                                                 ings of the lower-ranked condition. For return rate, we used
needed it, and to measure the effect of tutorial availability on
                                                                                 Pearson χ2 analyses to compare the percentages. The vari-
player engagement, we added another variable, availability
                                                                                 ance in our measures was very high, so a great deal of data
of help, with two possibilities: on-demand help, and no help.
                                                                                 was necessary to show statistical significance at the p = 0.05
We created a version of the game with on-demand help for
                                                                                 level.
each of the previously described conditions.
There are many ways to give tutorial information on-demand.                      Tutorials were only justified in the most complex game
For this experiment, we added a help button to the main tool-                    To evaluate whether tutorials improved player engagement,
bar of each game. Since Foldit’s main toolbar can be closed,                     we compared the version of the game with no tutorials
another floating help button was added near the score dis-                       to the versions with context-sensitive tutorials and context-
play. Clicking on this help button opened the same manual                        insensitive tutorials. The “presence” section of Table 3 shows
used in the context-insensitive tutorial condition described                     the results for these comparisons.
previously. The manual included only the concepts which the                      We expected to find that including tutorials, either in-context
player could have encountered up to that point. If later lev-                    or out-of-context, would lead to higher engagement than pro-
els introduced new concepts, the help manual did not include
                                                                                 viding no instruction at all. In Foldit, this was the case. Foldit
those concepts until the user reached that level. In order to
                                                                                 players with context-sensitive tutorials played 75% more lev-
make the help provided as context-sensitive as possible, the
                                                                                 els and 29% longer than those with no tutorials. Players with
manual automatically opened to the page that most closely re-                    context-insensitive tutorials played 25% more levels and 12%
lated to the level the user is currently playing. If desired, the                longer than with no tutorials. We found no significant effects
player could view all of the pages of the help screen with the                   for return rate.
“next” and “back” buttons. The player could close the help
screen and return to the current level by clicking on a “close”                  However, tutorials were not as effective in the other two
button. When activated, the help screen covered the main in-                     games. We found no significant effects for either compari-
teraction area for each game and restricted gameplay until the                   son in Refraction. In Hello Worlds, we found no significant
player closed the tutorial.
The Impact of Tutorials on Games of Varying Complexity
Experimental variable       Game             Condition         Player Count        Time Played          Levels Completed           Return Rate
                                           Context sensitive          1242       660s      p < 0.001     7      p < 0.001      19.16%      p = 0.408
                                              No Tutorials            1210       510s     Z = −5.070     4     Z = −10.982     20.50%      χ2 = 0.686
                               Foldit
                                          Context insensitive         1147       570s      p = 0.010      5     p < 0.001      19.62%      p = 0.594
                                              No Tutorials            1210       510s     Z = −2.592     4      Z = 4.353      20.50%      χ2 = 0.284
                                           Context sensitive          1634       990s      p = 0.437     15     p = 0.294      28.46%      p = 0.925
                                              No Tutorials            1678      1050s Z = −0.778 16            Z = −1.049      28.61%      χ2 = 0.009
         Presence            Refraction
                                          Context insensitive         1687      1020s      p = 0.483     15     p = 0.272      28.45%      p = 0.922
                                              No Tutorials            1678      1050s      Z = 0.702     16     Z = 1.099      28.61%      χ2 = 0.010
                                           Context sensitive          2817       750s      p = 0.697     10     p = 0.537      17.96%      p < 0.001
                                              No Tutorials            2815       720s     Z = −0.389     10    Z = −0.618      21.60% χ2 = 11.733
                           Hello Worlds
                                          Context insensitive         2754       690s      p = 0.572     10     p = 0.770      20.59%      p = 0.356
                                              No Tutorials            2815       720s     Z = −0.565     10     Z = 0.292      21.60%      χ2 = 0.854
                                           Context sensitive          1242       660s      p = 0.014     7      p < 0.001      19.16%      p = 0.779
                               Foldit
                                          Context insensitive         1147       570s     Z = −2.470      5    Z = −7.727      19.62%      χ2 = 0.079
                                           Context sensitive          1634       990s      p = 0.901     15     p = 0.994      28.46%      p = 0.998
    Context sensitivity      Refraction
                                          Context insensitive         1687      1020s Z = −0.124 15             Z = 0.008      28.45%      χ2 = 0.000
                                           Context sensitive          2817       750s      p = 0.348     10     p = 0.736      17.96%      p = 0.013
                           Hello Worlds
                                          Context insensitive         2754       690s     Z = −0.938     10    Z = −0.337      20.59%      χ2 = 6.175
                                                Blocking              1210       630s      p = 0.454     7      p = 0.007      18.93%      p = 0.881
                               Foldit
                                             Non-blocking             1242       660s     Z = −0.749     7      Z = 2.697      19.16%      χ2 = 0.022
                                                Blocking              1634       930s      p = 0.142     14     p = 0.139      26.37%      p = 0.187
         Freedom             Refraction
                                             Non-blocking             1551       990s     Z = −1.467     15    Z = −1.477      28.46%      χ2 = 1.743
                                                Blocking              2729       723s      p = 0.740      9     p = 0.376      18.94%      p = 0.346
                           Hello Worlds
                                             Non-blocking             2817       750s     Z = −0.332     10    Z = −0.886      17.96%      χ2 = 0.889
                                             Help (aggreg.)           4939       600s      p = 0.536      5     p = 0.208      18.34%      p = 0.130
                                           No help (aggreg.)          4809       600s     Z = −0.619      5    Z = −1.258      19.55%      χ2 = 2.297
                               Foldit
                                               Help Only              1238       570s      p = 0.036     4      p = 0.001      18.98%      p = 0.347
                                              No Tutorials            1210       510s     Z = −2.101     4     Z = −3.197      20.50%      χ2 = 0.885
                                             Help (aggreg.)           6608       960s      p = 0.806     15     p = 0.515      28.50%      p = 0.528
                                           No help (aggreg.)          6550       990s      Z = 0.245     15     Z = 0.651      28.00%      χ2 = 0.399
    Availability of help     Refraction
                                               Help Only              1678       900s      p = 0.031     14     p = 0.013      27.43%      p = 0.451
                                              No Tutorials            1655      1050s Z = −2.161 16            Z = −2.496      28.61%      χ2 = 0.569
                                             Help (aggreg.)          11302       720s      p = 0.626     10     p = 0.972      18.56%      p = 0.021
                                           No help (aggreg.)         11115       720s      Z = 0.487     10     Z = 0.035      19.78%      χ2 = 5.314
                           Hello Worlds
                                               Help Only              2864       754s      p = 0.434     11     p = 0.190      18.44%      p = 0.003
                                              No Tutorials            2815       720s     Z = −0.782     10    Z = −1.310      21.60%      χ2 = 8.875
Table 3. Summary of data gathered during the experiment, organized around our four experimental variables of presence, context-sensitivity, freedom,
and availability of help on-demand. Statistically significant results are shown in blue italics. Our results show that tutorials only had positive effects
on player behavior in Foldit, that context-sensitive tutorials outperformed context-insensitive tutorials in Foldit, that restricting player freedom did
not improve engagement in any game, and that providing access to help on-demand was beneficial in Foldit, ineffective in Hello Worlds, and actually
harmful in Refraction.

effects for levels completed and time played, and we found                      Context-sensitivity can improve engagement
that 3.5% fewer players returned with context-sensitive tuto-                   We next evaluated the importance of presenting tutorial in-
rials than with no tutorials.                                                   formation in context. We expected that presenting informa-
                                                                                tion as closely as possible to when the player needed it would
In contrast to our expectations, our results show that tutori-
                                                                                be better than presenting information out of context. The
als are only useful for improving player engagement in some
                                                                                results for the comparison of context-sensitive and context-
games. It may be the case that tutorial value depends on
                                                                                insensitive tutorials are shown in the “context-sensitivity”
game complexity. Refraction and Hello Worlds are typical
                                                                                section of Table 3.
of their genres and have intuitive interfaces. Foldit, on the
other hand, is unconventional, complex, and requires deep                       Context-sensitivity improved player engagement in Foldit.
strategy and spatial insight. Another possible explanation is                   Foldit players played 40% more levels and 16% longer
that Foldit players are more patient. Since the barrier to play                 with context-sensitive tutorials than context-insensitive tuto-
a downloadable game like Foldit is higher than for online ca-                   rials. However, we did not find positive effects for context-
sual games like Refraction and Hello Worlds, players who                        sensitivity in the other two games. We found no significant
are willing to overcome this barrier may be willing to devote                   effects for Refraction. In Hello Worlds, we found that the
more time reading the tutorials. Further work is necessary                      return rate was about 2% less with context-sensitive tutorials
to know the exact reasons for this effect. Regardless, our re-                  than with context-insensitive tutorials.
sults show that tutorials may only be worthwhile for complex
games, such as Foldit, and may not be worth the investment                      Therefore, the importance of context-sensitivity in a partic-
of time and resources in games with mechanics that are more                     ular game depends on whether tutorials positively impact
easily discovered through experimentation, such as Refrac-                      player engagement in that game. For complex games where
tion and Hello Worlds.                                                          tutorials are beneficial, such as Foldit, presenting the infor-
                                                                                mation in context seems beneficial. However, for games in
which tutorial presence does not have an impact, it does not        access to the help button in Refraction ever used it. One pos-
matter whether or not the information is presented in context.      sible explanation is that the knowledge that there is help avail-
                                                                    able discourages players from putting as much effort into the
                                                                    game. It is also possible that players will only click “help” if
Tutorial freedom did not affect player behavior                     they are already frustrated, and will quit even sooner if they
To evaluate whether restricting player freedom would im-            are unable to find the help they need. We can only specu-
prove tutorial effectiveness, we compared the versions with         late at this point, however, since multivariate testing does not
context-sensitive blocking tutorials to those with context-         tell us what players are thinking. Future studies are needed
sensitive nonblocking tutorials. The results of this compar-        to reproduce and understand this effect. Nevertheless, it is
ison are shown in the “freedom” section of Table 3.                 clear that players can respond unpredictably to tutorial im-
                                                                    plementations, and that tutorials can harm player retention
We expected that blocking tutorials would focus player atten-
                                                                    unexpectedly. This points to the importance of testing tuto-
tion and improve learning. In Foldit, we found a significant
                                                                    rial implementations to avoid unexpected negative effects.
effect for levels completed, but the median number of levels
completed remained constant at 7 levels. We found no signif-        CONCLUSION
icant effects for Refraction and Hello Worlds.
                                                                    Our examination of tutorials in three games of varying com-
Therefore, we found no evidence supporting the practice of          plexity showed that tutorial effectiveness depends on the
restricting player freedom in order to focus player attention       complexity of the game. Tutorials had the greatest value
on target interface objects. This result may further reinforce      in the most unconventional and complex game, Foldit, in-
our conclusion that players learn how to use the interface pri-     creasing play time by as much as 29%. However, tutorials
marily through experimentation, and that forcing the player to      had surprisingly little impact on player behavior in Refrac-
perform a series of actions may not be effective. It may also       tion and Hello Worlds, which are less complex, more similar
be the case that players do not like having their freedom re-       to other games, and easier to learn through experimentation.
stricted and that this cancels out any positive effects on learn-   Our results suggest that tutorials may not be effective for such
ing. Further examination is necessary to understand whether         games, and that designers should consider the complexity and
restricting player freedom improves player engagement and           discoverability of game mechanics when deciding whether to
learning in some cases.                                             invest resources in tutorials. It is unlikely that a single ap-
                                                                    proach will work for tutorial design in all games.

On-demand help harmed and helped player retention                   Since players seem to learn more from exploring than from
We next evaluated whether player engagement would be im-            reading text, we believe that it is important to design early
proved by providing additional help that players could access       levels in a way that maximizes a player’s ability to exper-
when needed. We expected that providing additional help in          iment and discover game mechanics. A key question that
this way would be beneficial. The results for the following         arises is how to facilitate this experimentation while ensur-
comparisons are shown in the “availability of help” section         ing that the player learns how to play and does not become
of Table 3.                                                         frustrated. We found little evidence to suggest that restrict-
                                                                    ing player freedom to focus attention on a particular interface
To look for large-scale effects of providing on-demand help         object or game mechanic is beneficial. Although it may be
in general, we first aggregated the four conditions with a help     tempting to provide help on-demand, we found that adding a
button and the four conditions without a help button and com-       help button was only effective in Foldit, and actually reduced
pared them against each other. We found only one significant        player progress by 12% and play time by 15% in Refrac-
effect across all three games, which was a 1.2% increase in         tion. Future work is needed to understand how to break down
return rate for Hello Worlds.                                       a complex game into smaller “chunks” that can be learned
                                                                    through exploration, how to detect when a player is confused
We then assumed that the effect of providing help on-demand
                                                                    or frustrated, and how to intervene, if necessary, in a way that
would be strongest when no other instructions were given.
                                                                    causes learning without negatively impacting engagement.
Therefore, we compared the version with a help button and
no other tutorials with the version with no help button and         One of the drawbacks of our methodology is that we do not
no tutorials. In Foldit, providing a help button in this case       know demographic information about our players. It is likely
increased engagement. Players played about 12% longer with          that each game attracted different kinds of players and that
a help button than without. We found a significant effect on        self-selection contributed to the differences in player behavior
the number of levels completed but this effect did not change       that we observed. However, we believe that it is most useful
the median. However, we found negative effects in the other         to tailor tutorials for the players who do self-select to play
two games. In Hello Worlds, players returned about 3% less          each game because they are the players who are likely to play
frequently with on-demand help than without. In Refraction,         the most.
we found that players with on-demand help completed 12%
fewer levels than those without access to help, and played for      Since we only tested tutorials in Refraction, Hello Worlds,
15% less time.                                                      and Foldit, we cannot know for sure how these results gener-
                                                                    alize to other games and genres. In particular, our games are
The negative impact of the help button, particularly in Re-         not representative of commercial games, and as a result we
fraction, is difficult to explain. Only 31% of players that had     can draw no conclusions about the effectiveness of tutorials
in games that players must purchase. Further research with a                       structures with a multiplayer online game. Nature, 466(7307):756–760,
wider variety of games is required to determine whether our                        August 2010.
results will apply to commercial games and games of differ-                     6. D. K. Farkas. The role of balloon help. ACM SIGDOC Asterisk Journal
ent genres.                                                                        of Computer Documentation, 17, 1993.
                                                                                7. J. P. Gee. Learning by design: Games as learning machines. Interactive
Furthermore, future work is necessary to understand how tu-                        Educational Multimedia, 8:15–23, 2004.
torial presentation affects player behavior. There are many                     8. Google Website Optimizer. http://www.google.com/websiteoptimizer/.
ways to design tutorials, and we only experimented with ba-
                                                                                9. F. Grabler, M. Agrawala, W. Li, M. Dontcheva, and T. Igarashi.
sic designs. Although video game tutorials frequently include                      Generating photo manipulation tutorials by demonstration. In ACM
audio, animations, and videos, we only examined tutorials                          SIGGRAPH 2009, New York, NY, USA, 2009. ACM.
containing pictures and text. Another question is whether de-                  10. T. Grossman and G. Fitzmaurice. Toolclips: An investigation of
signers should integrate tutorials into the main game progres-                     contextual video assistance for functionality understanding. In CHI ’10
sion or provide them in a separate “tutorial mode” that can                        Proceedings of the 28th international conference on Human factors in
be avoided entirely, as suggested by [1] and utilized in highly                    computing systems, New York, NY, USA, 2010. ACM.
successful games such as Deus Ex (Ion Storm Inc. 2000). Fu-                    11. T. Grossman, G. Fitzmaurice, and R. Attar. A survey of software
ture experiments with a greater variety of tutorial styles will                    learnability: Metrics, methodologies and guidelines. In CHI ’09:
                                                                                   Proceedings of the 27th international conference on Human factors in
improve our understanding of how to design effective tutori-                       computing systems, New York, NY, USA, 2009. ACM.
als.
                                                                               12. S. M. Harrison. A comparison of still, animated, or nonillustrated
Our results point to the importance of analytics and large-                        on-line help with written or spoken instructions in a graphical user
scale analysis of player behavior because these effects would                      interface. In CHI ’95 Proceedings of the SIGCHI conference on Human
                                                                                   factors in computing systems, New York, NY, USA, 1995. ACM.
be difficult to discover without large amounts of data. We
believe that this methodology will continue to provide valu-                   13. J. Huang and M. B. Twidale. Graphstract: Minimal graphical help for
                                                                                   computers. In UIST ’07 Proceedings of the 20th annual ACM
able insights that improve our understanding of good game                          symposium on User interface software and technology, New York, NY,
design. Such experiments are important because player be-                          USA, 2007. ACM.
havior is often counterintuitive; each of our four hypotheses                  14. C. Kelleher and R. Pausch. Stencils-based tutorials: Design and
turned out to be either incorrect or incomplete. Furthermore,                      evaluation. In CHI ’05 Proceedings of the SIGCHI conference on
we believe that the ability of games to attract large numbers of                   Human factors in computing systems, New York, NY, USA, 2005.
participants, allowing researchers to perform tests with many                      ACM.
experimental variables, will make them an important mecha-                     15. K. Knabe. Apple guide: A case study in user-aided design of online
nism for future HCI research.                                                      help. In CHI ’95 Conference companion on Human factors in
                                                                                   computing systems, New York, NY, USA, 1995. ACM.
ACKNOWLEDGMENTS                                                                16. R. Kohavi, R. M. Henne, and D. Sommerfield. Practical guide to
We would like to acknowledge the members of the Refrac-                            controlled experiments on the web: listen to your customers not to the
                                                                                   hippo. In KDD ’07: Proceedings of the 13th ACM SIGKDD
tion, Hello Worlds, and Foldit teams for developing the games                      international conference on Knowledge discovery and data mining,
and making this work possible. We also thank Anthony                               pages 959–967, New York, NY, USA, 2007. ACM.
Pecorella and August Brown of Kongregate for promoting                         17. T. Lau, L. Bergman, V. Castelli, and D. Oblinger. Sheepdog: Learning
our games and helping us gather data. This work was sup-                           procedures for technical support. In IUI ’04 Proceedings of the 9th
ported by the University of Washington Center for Game Sci-                        international conference on Intelligent user interfaces, New York, NY,
ence, DARPA grant FA8750-11-2-0102, the Bill and Melinda                           USA, 2004. ACM.
Gates Foundation, NSF grant IIS0811902, two NSF Graduate                       18. J. Lave and E. Wenger. Situated learning: Legitimate peripheral
Fellowships, Adobe, and Intel.                                                     participation. Cambridge University Press, Cambridge, England, 1991.
                                                                               19. G. Linden. Early Amazon: Shopping cart recommendations, 2006.
REFERENCES                                                                         http://glinden.blogspot.com/2006/04/early-amazon-shopping-cart.html.
 1. E. Adams. The designer’s notebook: Eight ways to make a bad tutorial.
                                                                               20. J. Nielsen. Usability Engineering. Morgan Kaufmann, San Francisco,
    Gamasutra, 2011.
                                                                                   CA, USA, 1993.
 2. E. Andersen, Y.-E. Liu, R. Snider, R. Szeto, S. Cooper, and Z. Popović.
    On the harmfulness of secondary game objectives. In FDG ’11:               21. S. Palmiter and J. Elkerton. An evaluations of demonstrations for
    Proceedings of the Sixth International Conference on the Foundations           learning computer-based tasks. In CHI ’91 Proceedings of the SIGCHI
    of Digital Games, New York, NY, USA, 2011. ACM.                                conference on Human factors in computing systems: Reaching through
                                                                                   technology, New York, NY, USA, 1991. ACM.
 3. E. Andersen, Y.-E. Liu, R. Snider, R. Szeto, and Z. Popović. Placing a
    value on aesthetics in online casual games. In CHI ’11: Proceedings of     22. S. G. Ray. Tutorials: Learning to play. Gamasutra, 2010.
    the SIGCHI conference on Human factors in computing systems, New           23. B. Shneiderman. Designing the User Interface: Strategies for Effective
    York, NY, USA, 2011. ACM.                                                      Human-Computer Interaction. Addison-Wesley Longman Publishing
 4. L. Bergman, V. Castelli, T. Lau, and D. Oblinger. Docwizards: A                Co., Inc., Boston, MA, USA, 1986.
    system for authoring follow-me documentation wizards. In UIST ’05
                                                                               24. J. Xiao, R. Catrambone, and J. Stasko. Be quiet? evaluating proactive
    Proceedings of the 18th annual ACM symposium on User interface
                                                                                   and reactive user interface assistants. Technical report, Georgia Institute
    software and technology, New York, NY, USA, 2005. ACM.
                                                                                   of Technology, Atlanta, GA, USA, 2003.
 5. S. Cooper, F. Khatib, A. Treuille, J. Barbero, J. Lee, M. Beenen,
    A. Leaver-Fay, D. Baker, Z. Popović, and F. Players. Predicting protein
You can also read